Presentation is loading. Please wait.

Presentation is loading. Please wait.

Group-Collective Communicator Creation Ticket #286 Non-Collective Communicator Creation in MPI. Dinan, et al., Euro MPI ‘11.

Similar presentations


Presentation on theme: "Group-Collective Communicator Creation Ticket #286 Non-Collective Communicator Creation in MPI. Dinan, et al., Euro MPI ‘11."— Presentation transcript:

1 Group-Collective Communicator Creation Ticket #286 Non-Collective Communicator Creation in MPI. Dinan, et al., Euro MPI ‘11.

2 Non-Collective Communicator Creation  Create a communicator collectively only on new members  Global Arrays process groups –Past: collectives using MPI Send/Recv  Overhead reduction –Multi-level parallelism –Small communicators when parent is large  Recovery from failures –Not all ranks in parent can participate  Load balancing –Reassign processes from idle groups to active groups 2 

3 Non-Collective Communicator Creation Algorithm 3 MPI_COMM_SELF (intracommunicator) MPI_Intercomm_create (intercommunicator) MPI_Intercomm_merge ✓ 01234567 01234567 01234567 01234567

4 Non-Collective Algorithm in Detail RR 4 Translate group ranks to ordered list of ranks on parent communicator Calculate my group ID Left group Right group

5 Evaluation: Microbenchmark 5 O(log 2 g) O(log p)

6 048 1212… 3232 3636 40404… MCMC Benchmark Kernel  Simple MPI-only benchmark  “Walker” groups –Initial group size: G = 4  Simple work distribution –Samples: S = (leader % 32) * 10,000 –Sample time: t s = 100 ms / |G| –Weak scaling benchmark  Load balancing –Join right when idle –Asynchronous: Check for merge request after each work unit –Synchronous: Collectively regroup every i samples 6

7 Evaluation: Load Balancing Benchmark Results 7

8 Ticket #286: Group-collective communicator creation MPI_Comm_create_group(comm, group, tag, newcomm) IN commcommunicator (handle) IN groupGroup, which is a subset of the group of comm (handle) IN tag"safe" tag (integer) OUT newcommnew communicator (handle) int MPI_Comm_create_group(MPI_Comm comm, MPI_Group group, int tag, MPI_Comm *newcomm) MPI_COMM_CREATE(COMM, GROUP, TAG, NEWCOMM, IERROR) INTEGER COMM, GROUP, KEY, NEWCOMM, IERROR 8

9 MPI_COMM_CREATE_GROUP provides the same semantics as MPI_COMM_CREATE, however comm must be an intracommunicator and this routine need only be invoked by processes in the input group, group. Like MPI_COMM_CREATE, if a process calling MPI_COMM_CREATE_GROUP is not a member of given group it returns MPI_COMM_NULL. If different groups are provided by different callees, the calls will be interpreted as separate group-collective communicator construction operations; each call requires all members of the group to invoke this routine. This results in the same behavior as MPI_COMM_CREATE with different group arguments. Advice to users: Group-collective creation of an intercommunicator can be achieved by creating the local communicator using MPI_COMM_CREATE_GROUP and using this as the input to MPI_INTERCOMM_CREATE. This call may use point-to-point communication with communicator comm, and tag tag between processes in group. Thus, care must be taken that there be no pending communication on comm that could interfere with this communication. Likewise, if a given process performs multiple subset-collective communicator creation calls, the user must be careful to order calls and use different tag arguments to ensure that calls match correctly. Advice to users: MPI_COMM_CREATE may provide lower overhead because it can take advantage of collective communication on comm when constructing newcomm. Proceed with proposal? Yes ( 16 ), No ( 0 ), Abstain ( 1 ) 9


Download ppt "Group-Collective Communicator Creation Ticket #286 Non-Collective Communicator Creation in MPI. Dinan, et al., Euro MPI ‘11."

Similar presentations


Ads by Google